Multi-agent AI platform mines thousands of clinical trials to pinpoint features of successful drug targets

Scientists report today in Science a multi-agent artificial intelligence system that can analyze and integrate biomedical data across scales to inform drug development decisions. The work, led by senior author James Zou at Stanford University, with first author Harrison G. Zhang, demonstrates that the platform—dubbed the Virtual Biotech—can uncover patterns associated with clinical trial success, propose therapeutic strategies, and retrospectively explain trial failures.

The Virtual Biotech is structured like a drug discovery company: a virtual Chief Scientific Officer orchestrates specialized AI agents—each focused on areas such as genetics, single-cell biology, safety, or clinical trials—that retrieve and analyze primary data. The system has access to over 100 analytical tools spanning 78,726 targets, 39,530 diseases, 14.5 million protein interactions, and more than 100 million single-cell profiles.

In its first case study, the platform deployed more than 37,000 agents to annotate outcomes from nearly 56,000 clinical trials. The analysis revealed that drugs targeting genes with cell-type-specific expression were 48% more likely to reach the market and had 32% fewer adverse events compared with broadly expressed targets. The finding held even after adjusting for genetic evidence, suggesting that single-cell transcriptomic data provide independent information for target prioritization.

A second case study focused on the immune checkpoint protein B7-H3 in lung cancer. The Virtual Biotech integrated single-cell, spatial, and clinical data to propose that B7-H3 is overexpressed in cancer-associated fibroblasts rather than tumor cells alone, and that this expression correlates with an immune-excluded microenvironment. Based on these insights, the system recommended an antibody–drug conjugate (ADC) as the preferred modality—a strategy later supported by a regulatory breakthrough designation for a B7-H3–targeted ADC.

In a third analysis, the system examined a terminated Phase II trial for an ulcerative colitis drug targeting OSMR. By analyzing single-cell data from treatment-refractory patients, the agents found that while OSMR was upregulated, downstream JAK-STAT signaling was driven by multiple redundant receptors, suggesting that blocking OSMR alone might be insufficient. The system then derived a composite biomarker that outperformed single-gene measures in predicting non-response across independent cohorts.

The authors emphasize that the Virtual Biotech is a decision-support tool, not a replacement for experimentation. Its conclusions depend on available data and are strongest for well-characterized human diseases. The analyses are observational and require validation. Still, the platform illustrates how coordinated AI systems could make drug discovery more transparent, scalable, and hypothesis-driven.

As agentic systems mature, they may enable a model where hypotheses are continuously generated, stress-tested against diverse evidence, and refined in partnership with human scientists—potentially accelerating the slow and costly process of bringing new therapies to patients.

---

Paper reference:
Harrison G. Zhang, Peter Eckmann, Jiacheng Miao, Andrew B. Mahon, James Zou. The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development. Science (2026). DOI: 10.1126/science.aeg6779